Ple Table

class PleTable

Immutable embedding row layout. Opening validates metadata, not payload digests.

Public Types

enum class DType

Values:

enumerator kFP8

enumerator kFP16

enumerator kBF16

enumerator kFP32

Public Functions

RowLocation locate(int64_t row) const

Translate a valid global row into an exact file range without reading payload bytes.

inline int64_t rowCount() const noexcept

inline int64_t rowWidth() const noexcept

inline int64_t rowBytes() const noexcept

inline DType dtype() const noexcept

inline DType sourceDtype() const noexcept

inline int64_t sourceRowBytes() const noexcept

inline double sourceScale() const noexcept

inline double scale() const noexcept

inline std::string const &tableId() const noexcept

inline std::vector<Shard> const &shards() const noexcept

Public Static Functions

static PleTable fromSource(

Source const &source,

int64_t rowCount,

int64_t rowWidth,

std::filesystem::path const &checkpointDir,

std::filesystem::path const &dataFile = {},

DType poolDtype = DType::kFP16,

double poolScale = 1.0

)

Original checkpoint tensor contract. An explicit location describes indexed PyTorch payloads. A table split across tensors sets shards instead of tensorKey, in row order. struct Source { std::string tensorKey; DType dtype{DType::kFP16}; int64_t bytes{}; double scale{1.0}; std::vector<Sample> samples; std::filesystem::path file; int64_t offset{}; std::vector<ShardSource> shards; };

struct Shard { std::filesystem::path path; int64_t rowBegin{}; int64_t rowCount{}; int64_t byteOffset{}; int64_t byteLength{}; };

struct RowLocation { size_t shardIndex{}; int64_t byteOffset{}; };

! Expected identity and scale come from the engine’s bound artifact metadata. static PleTable open(std::filesystem::path const& manifest, std::string const& expectedTableId, double expectedScale, std::filesystem::path const& dataFile = {});

! Resolve the source tensors from a checkpoint or explicit file and validate their metadata and ! bounded samples. poolDtype/poolScale describe the rows the pool will hold, which differ from the source when the fill converts; the defaults are the cast-to-F16 contract.

struct Sample

Public Members

int64_t offset = {}

std::string data

struct ShardSource

One checkpoint tensor holding a contiguous row range. Sample offsets are relative to it.

Public Members

std::string tensorKey

int64_t rows = {}

int64_t bytes = {}

std::vector<Sample> samples

std::filesystem::path file

int64_t offset = {}

struct Sample

Public Members

int64_t offset = {}

std::string data

struct ShardSource

One checkpoint tensor holding a contiguous row range. Sample offsets are relative to it.

Public Members

std::string tensorKey

int64_t rows = {}

int64_t bytes = {}

std::vector<Sample> samples

std::filesystem::path file

int64_t offset = {}